The AI Company Simulating the Entire Economy | Simile Co-founder & CEO, Joon Sung Park

Summary

• Park’s investable thesis, stated twice and worth pricing in: “for AI companies of this generation, you need to have an interesting data strategy that’s going to be defensible.” Simile’s data strategy goes beyond web data — which records what people say — to behavior, transaction, and above all randomized-control-trial data, because “no one really cares about prediction… What people actually care about is they want to shape the future,” and that requires causal counterfactuals. Asked where he’d invest, he applies the same test: data nobody else can access or collect — naming robotics and the inference/chip layer.
• Some enterprise deals closed within 3 months. Park planned to spend until end-2026 warming up the market; instead some of the largest enterprise customers closed “at a lightning speed for enterprise” because the pain of slow, expensive experimentation was “way more acute than I could have imagined.” The killer demo: rerunning a consulting firm’s study on the first call — “we predicted the outcome of studies that took 3 to 6 months, but just within 2 minutes.”
• Validation number to anchor on: 85%. After the Smallville demo drew inbound from Fortune 500 boards, the team spent a year proving models predict people’s behaviors and attitudes “85% as accurately as people replicate their own” (published end-2024) — the work Park credits with starting the synthetic-panels field, which he says will outgrow the current human-panel market within 3 years since only ~5% of those ideas ever get answered.
• The pricing endgame is extreme: “in about 2-3 years, we’re running a single simulation session that’s going to take 10, 20 million dollars to run… but it’s going to be so valuable that people will pay $100 million for it” — simulation as the next frontier of token-maxed inference, sold to the largest enterprises and governments. Today’s production model already runs at ~1/100th its original cost.
• On Kalshi/Polymarket and markets generally: Simile’s differentiation is “not just what’s going to happen, but how it’s going to happen and why” — showing the steps so customers can prevent or encourage the outcome. Quants have already joined the firm, “maybe Simile will actually own a small hedge fund down the line,” and with some form of AGI and a perfect simulator, Park thinks assumptions we hold about the world will change — “certainly, one of these could actually be the stock market.” The deeper assumption he says may no longer hold is that everyone’s perspective is impossible to obtain — replaced by “a representational layer of our society.”
• $300M raised in ~6 months: a $100M round ~5 months ago, then a $200M insider preempt — Shardul at Index has “never seen this kind of traction,” and Greenoaks had independently mapped the market and closed in days without a process. Park said there was certainly a consideration of whether they needed the money; the raise logic was “the money does take compute” — you can’t control research outcomes, only inputs.
• Team-building doctrine worth stealing: hire people who were “the common denominator of success” at every stage of their life, and who hold “two superpowers that’s not supposed to coexist” — his co-founder Laney is “short-term paranoid, long-term religious,” a balance requiring “somebody who is broken in some ways.” Harry’s addendum: founders who wish they’d worried less have it backwards — the paranoia is what made it work out.

Deep dive

1. Smallville: the Valentine’s Day simulation that started the field

• The founding artifact is Park’s 2023 experiment: a game town of 25 NPCs powered by GPT-3.5 text-davinci — pre-ChatGPT — paired with memory, planning, and reflection, “really the first times that those concepts came out to be an explicit part of the architecture” of agentic workflows. Set the day before Valentine’s Day, the agents self-organized: planning parties, decorating the cafe, remembering each other.
• The memory solution began embarrassingly simply — “we’ll put everything in markdown text file. That was it” — but the key addition was reflection: a scheduled “shower thought” where the agent interrogates its own logs (“Why did you get omelet so often this week? Were you busy? Do you like omelet?”) and formulates ideas above ground truth. “This actually shapes who they are as a person” — agents acquire personality and a point of view.

2. Not a smarter model — a foundation model of human error

• Park’s positioning relative to frontier labs: they build “super rational intelligent machines that are good at coding, natural sciences and mathematics. Simile doesn’t really care about any of those. What we care about is if we have a person make a mistake in this context, we want our models to make the same kind of mistake… biased in the same way humans are” — the “subjective half of their brain”: values, preferences, taste.
• Harry’s blunt challenge — do you just want to be a next-generation Qualtrics? Park’s answer: surveys and interviews are the tooling layer; simulation is “the most generalizable model of people… at scale,” extending to simulating an entire product launch, then “wicked problems” — climate-change coordination equilibria, even “in what conditions does a democracy fail?”
• On misuse (Harry raised elections): simulation is one of sci-fi’s “twin pillars” alongside AGI, and its potential for abuse “is quite real.” The stated North Star: “representation at scale” — putting the viewpoints of people absent from decision-making rooms into every decision.

3. The data thesis: causality over correlation, RCTs over web data

• The episode’s spine: “my fundamental thesis here is for AI companies of this generation, you need to have an interesting data strategy that’s going to be defensible.” Web data is what people said, not did; Simile collects transaction and observational data, but Park’s “personal hot take” is that observational data primarily helps create correlation — “good for prediction task. But… no one really cares about prediction… unless you’re trying to predict the stock market.”
• The Starbucks example carries the argument: telling them Frappuccino sales will tank in two quarters just gets “What do we do about them? That’s terrible.” Customers want to shape the future, which requires causal mechanisms and counterfactuals — so RCTs and A/B tests (“imagine people have done this versus that”) are the core training asset.
• Sourcing differs from the labs too: not expert programmers but “people like us… everyday people,” with representativeness a key concern, opened with “Tell us the story of your life.” On scale: 1,000 people gets statistical significance for a narrow population, but customers filter on the fly — “that means we want to represent the entire population.”

4. The reward function: “the world is our ground truth”

• Harry’s AlphaGo analogy prompted Park’s sharper version. Coding agents improved fast because accept/reject gave a crisp reward; simulation looks unrewardable since predictions live in the future — but “the world is our ground truth. Every single day, we can be generating tens of thousands of hypotheses… a month goes by, we generated a million hypotheses, X percentage of them came true. This is the best way to learn about the world.”
• The flywheel compounds for early partners — “absolutely, yes” — and cost curves bend hard: the model now in production “used to cost about 100 times more to run than it does now,” same reward model and philosophy. Heavy compute buys the “initial point of view”; efficiency follows.

5. Product-market fit arrived years early — and closed in weeks

• PMF announced itself when Fortune 500 board members and C-suites reached out after seeing the Smallville demo at Stanford. The team spent a year validating, showing models predict behaviors and attitudes “85% as accurately as people replicate their own” — published end-2024, “that’s really what started the field around synthetic panels.”
• Park expected a 1-2 year warm-up with aggressive go-to-market toward end-2026. Instead — and this “truly made me change my perspective on corporate America” — some of the largest enterprise customers closed within 3 months, because the pain of slow experimentation “was way more acute than I could have imagined.” He highlights CVS and its VP of Insights, Shree, as an especially forward-looking counterpart.
• The first-call sell: customers handed over a finding from a large consulting firm and asked Simile to rerun it — “we predicted the outcome of studies that took 3 to 6 months, but just within 2 minutes.”
• Will synthetic panels beat human panels in 3 years? Yes, because the ceiling rises: only ~5% of the ideas companies and scientists want tested ever get answered — “a lot of the decisions that we make as a society, we base on our gut instinct.” Enterprise market research is explicitly a wedge, per Pat Hanrahan’s Tableau advice: “the best way to get feedback is to actually ask people to pay you.”

6. The economics: $100M simulation sessions and a possible hedge fund

• On Harry’s value-extraction question (moving billions for CVS while charging a million), Park’s answer is both, with prevention a huge value case: avoiding “a total disaster… that would have cost us half a billion dollars” is “a true painkiller.” Complex simulations — full downstream implications, US-wide segmentation — cost more but carry the highest ROI.
• The scale call of the episode: “I think there’s a world in which in about 2 3 years, we’re running a single simulation session that’s going to take 10, 20 million dollars to run… but it’s going to be so valuable that people will pay 100 million dollars for it” — for the largest enterprises and governments, simulation as the next frontier of token-maxed inference.
• Versus Kalshi/Polymarket: overlap in caring about the future, but “we are a company that is not just interested in what’s going to happen, but more on how it’s going to happen and why” — showing every step the ecosystem takes so customers can prevent or encourage it.
• Harry’s hedge-fund provocation landed: quants have already joined Simile, “maybe Simile will actually own a small hedge fund down the line.” And could markets become uninvestable? Given some form of AGI and some form of perfect simulator, Park thinks many assumptions we hold about the world will change. Certainly, one of these could actually be the stock market. The deeper assumption he says may no longer hold: that it’s impossible to get everyone’s perspective — replaced by “a representational layer of our society.”

7. $300M in six months: the preempt mechanics

• Simile raised $100M ~5 months ago, then was preempted by insiders: Shardul at Index, who led the prior round, has “quite never seen this kind of traction, this kind of pull.” Park called his top-of-list outside team — Greenoaks — who had already been mapping the market: “we’re not running a process, so if you’d be interested in joining, we have a few days.” The $200M round closed, bringing totals to $300M over ~6 months (seed led by Mike Volpi).
• They considered whether they needed the money; Park’s resolution was that “the money does take compute” — in research “you really cannot control the outcome… what you can control is the input and the process,” and capital meaningfully raises the input.
• His honest read on VCs, kept with its rough edges: “I was actually fairly skeptical what the roles of VCs actually were… I can’t still quite put my finger on it” — but the right ones become genuine mentors (Volpi introduced him to co-founder Laney). And timing lesson: seed, A, and the next round all landed within a year — “the market is always moving perhaps one step ahead of where you are.”
• On Harry’s hubris question: “there’s parts of market that is actually quite frothy. For sure.” His anchor is fundamentals — like OpenAI/Anthropic, the model-improvement rate and demand pull are mappable.

8. Hiring doctrine: common denominators and contradictory superpowers

• Park’s painter’s analogy (he was a figure painter): whoever the subject, “your subject sort of looks like the painters themselves” — and the best teams mirror their founder. Filter one: across every stage of a candidate’s life, “were they the common denominator” of success? Yes means extreme ownership and reinvention — co-founder Michael Bernstein went crowdsourcing → AI → generative agents, and “you could see that this is the person who led a lot of the success.”
• Filter two: “two superpowers that’s not supposed to coexist in one person.” Harry’s example, which Park endorsed: the handful of world-class CMOs who are “unbelievably data rigorous” and creatively artistic.
• The archetype in full, via co-founder Laney: daily she’s paranoid — “unless we put everything on our table today… we’ll lose” — but long-term “she’s religious,” believing “the world is stacked for her.” Balancing both “needs somebody who is broken in some ways.” Harry’s riff, worth keeping: successful founders’ stock answer “I wish I’d known it would all work out” is “the worst answer” — the paranoia drove the prep that produced the outcome.

9. The talent war, researcher-founders, and the quickfire calls

• On the researcher comp arms race: “absolutely” real — Park’s closest colleagues earn total comp “in the tens of millions,” and he’s upfront that Simile can’t match base “doesn’t matter how many hundreds of millions that you raised.” What wins them: vision and impact — “these are people who have literally seen OpenAI being the laughing stock… to becoming a nearly trillion-dollar business” within recent memory. His retention proof: in 6 years his core research team never left, and he convinced his own doctoral advisors — Bernstein and Percy, who “literally coined the term foundation model” — to join.
• His diligence test for academic founders, for investors wary of funding science projects: “are they married to a problem or are they married to impact?” Problem-married researchers often produce “not a good company”; impact-married ones hunt problems that reach people and generate revenue.
• Quickfire: most overheated — “new labs without a clear vision for how they’re going to impact the world… they will turn out to be interesting research project, but not a viable company.” Where he’d invest: the defensible-data test again — robotics (interesting, though hardly underinvested) and the inference and chip/hardware layer, where he’s “quite bullish” on one team recently out of stealth (likely Etched).
• The kindest thing: jobless in a Palo Alto garage with zero research background, he cold-messaged academics; Stanford theory professor Mary Wootters — same undergraduate college — gave him a full morning and the introductions that opened research to him. “I really didn’t think I deserved it, but that was the bet that they took.” Harry’s close: “never forget the first believer.”